If your threat model includes government agencies, using a third party like OpenAI that could be easily man in the middled by government agencies through purchasing OpenAI would seem counter to that goal.

It would be an interesting way to defeat stylometry techniques though.